Jiahuan Zhang, Shunwen Bai, Tianheng Wang, KaiWen Guo, Zijia Song, Hanqing WU, Guozheng Rao, Kai Han, Kaicheng Yu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Jiahuan Zhang, Tianheng Wang, Ziyi Huang, Yulong Wu, Hanqing Wu, DongbaiChen DongbaiChen, Linfeng Song, Yue Zhang, Guozheng Rao, Kaicheng Yu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The rapid development of social media platforms has accelerated the generation and spread of fake news. News on different platforms varies significantly in content and audience. It makes most existing fake news detection models, which rely on single-source datasets, struggle to perform well on news from other sources. Many social platforms also lack high-quality annotated data. To address this problem, we propose a novel cross-source multi-modal fake news detection model named CMFNThinker. CMFNThinker simulates human thinking patterns. It detects fake news across platforms in three stages: summarizing the news content, retrieving similar news posts and reasoning the truthfulness of the news. We conducted extensive experiments on multi-source datasets. The results show that our model outperforms state-of-the-art baseline models by at least 11.3% in macro F1 for cross-source fake news detection.

Complex Query Answering over incomplete knowledge graphs is a fundamental yet challenging task. Existing methods based on a pretrained knowledge graph embedding model have achieved good performance. However, they ignore logical rules. Logical rules, as part of the conceptual layer in knowledge graphs, contain rich background information that enhances logical reasoning and improves the performance of models. To address this problem, we propose a model that incorporates logical rules for complex query answering over knowledge graphs called R-CQA (Complex Query Answering with Rules). This model introduces implicit and explicit rule injection modules to the existing CQA models. The implicit rule injection module models logical rules as additional data for training and incorporates rule information into knowledge graph embedding model, which enhances inference. The explicit rule injection module uses logical rules to rewrite queries and constructs a query set for each original query, which avoids missing answers for inference. However, implicit rule injection does not really use logical rules for reasoning which decreases the interpretability and accuracy of logical rules. Explicit rule injection is influenced by the quantity and quality of logical rules. Therefore, we combine both methods to take advantage of their respective strengths. Experiments on 3 datasets demonstrate our model obtains state-of-the-art performance on complex query answering.

The proliferation of misinformation on the internet and social media platforms poses a significant challenge to public discourse integrity. Traditional Fact Verification methods rely on costly, domain-specific annotated datasets, limiting their adaptability. While Large Language Models (LLM) offer new possibilities, existing LLM-based techniques face critical limitations: (1) Dependence on simplistic pipelines: Lacking robustness to handle contradictory evidence or ambiguous claims. (2) Underutilization of multi-agent interactions: Restricting thorough claim assessments from multiple perspectives. To address these challenges, we propose a novel Fact - based Debate (FBD) framework, which combines retrieval-augmented generation (RAG) with iterative argumentation using structured multi-agent debate. On the one hand, the FBD framework is introduced to enhance the robustness and reliability of fact - checking. On the other hand, the FBD is designed with a three - stage pipeline for knowledge retrieval, acquisition, and refinement based on authoritative sources. Extensive experiments on five real - world datasets demonstrate that our proposed FBD outperforms existing methods, achieving an average relative improvement of approximately 2.36% compared to the optimal baseline.

Document-level event argument extraction aims to identify the event arguments and predict their roles in a document. There are many non-argument entities that play an essential role in understanding events in the document, which we call event-linking entities. However, recent work on document-level event argument extraction overlooks these entities and gets suboptimal performance. Moreover, a document usually contains multiple events, and they are interconnected with each other. Most recent work models each event in isolation without considering their interconnection. To tackle these problems, we propose a Generative Dual Representations Fusion network (GDRF) for document-level event argument extraction to introduce the significance of event-linking entities for the first time. GDRF retrieves event-linking entities and tags them in the document using the Event-linking Entity based Document Tagging module. To introduce contextual information from event-linking entities and capture the interconnection between multiple events, we propose a Dual Representations Fusion module to fuse original and tagged context representations with a fusion loss. Empirical results on the WIKIEVENTS dataset demonstrate that our model outperforms previous methods, achieving state-of-the-art performance.

Event detection is an eminent task in natural language processing, which aims at detecting event triggers in sentences and classifying them into specific event types. Some recent work has achieved good results in event detection tasks using syntactic dependency structures. However, the syntactic structure of parsing does not only bring benefits. It can also introduce noise or even misleading judgments. Moreover, most recent work focuses only on the contribution of syntactic information while ignoring the impact of semantic information. In this paper, we propose a novel Event Detection Model Based on the Fusion of Hierarchical Syntactic and Type Semantic Features (FHSTSF), which consists of a syntactic feature extractor for capturing hierarchical syntactic dependency features and a semantic feature extractor for extracting semantic features of event type labels. We conducted experiments on two public benchmark datasets, MAVEN and ACE-2005. The experimental results show that our model improves by 1.75% and 4.17% of the F1 Value on both benchmarks than the current best models, respectively.

Conversational Causal Emotion Entailment (C2E2) aims to identify the causes of a target emotion in a non-neutral conversation. Most models treat C2E2 as an independent utterance pair classification problem that ignores the contextual information. Furthermore, most recent works focus only on the contribution of utterance information while ignoring the impact of speaker and emotional information. To solve these problems, we propose a joint model with the contextual and speaker information for conversational causal emotion entailment. We introduce the temporal convolutional network structure to effectively capture contextual information and help the model better understand and analyze emotional changes in conversations. At the same time, a multi-feature interaction network is proposed to use multiple features of utterance to analyze the causes of the emotion, the network extracts location information from the position-aware graph and uses speaker and emotion information to help the model understand the causes behind emotion generation. The experimental results demonstrate that our method outperforms the baseline method and can infer the causes of different emotions in more complex contexts.

Abstract Event Detection (ED) is a crucial information extraction task that aims to identify the event triggers and classify them into predefined event types. However, most existing methods did not perform well when processing events with implicit triggers. And most methods considered ED as a sentence-level task, lacking effective context for event semantics. Moreover, how to maintain good performance under low resource conditions still needs further study. To address these problems, we propose a novel end-to-end ED model called DE3TC, which Detects Events with Effective Event Type Information and Context. We construct an event type-specific Clue to capture the interaction between event type name and trigger words, providing event type information for implicit triggers. For accessing the effective context of event semantics for sentence-level ED, we consider the correlations between types and select similar types’ descriptions as context. With contextualized representation from a contextual encoder, DE3TC learns the event type information for all events including implicit ones. And it performs sentence-level ED efficiently with effective contexts. The empirical results on ACE 2005 and MAVEN datasets show that: (i) DE3TC obtains state-of-the-art performance compared with previous methods. (ii) DE3TC is also excelled under low-resource conditions.

Pre-Trained Models (PTMs) can learn general knowledge representations and perform well in Natural Language Processing (NLP) tasks. For the Chinese language, several PTMs are developed, however, most existing methods concentrate on modern Chinese and are not ideal for processing classical Chinese due to the differences in grammars and semantics between these two forms. In this paper, in order to process two forms of Chinese uniformly, we propose a novel Classical and Modern Chinese pre-trained language model (CANCN-BERT), with the advantage of effectively processing both classical and modern Chinese, which is an extension of BERT. Form-aware pre-training tasks are elaborately designed to train our model, so as to better adapt it to classical and modern Chinese corpus. Moreover, we define a joint model, proposing dedicated optimization methods through different paths with the control of the switch mechanism. Our model merges characteristics of both classical and modern Chinese, which can adequately and efficiently enhance the representation ability for both forms. Extensive experiments show that our model outperforms baseline models on processing classical and modern Chinese and achieves significant and consistent improvements. Also, the results of ablation experiments demonstrate the effectiveness of each module.

Aspect-based sentiment analysis (ABSA) aims to determine the sentiment polarity of the specific aspect for a given sentence. Attention-based models are widely used in this task because they can extract semantic information between context words to make up for the deficiency of sequence models in semantic encoding. In order to enhance the extraction of high-quality semantic information, we propose a novel joint model with Second-Order Features and Matching Attention (SOMA) for aspect-based sentiment analysis. Firstly, we introduce the second-order statistics to extract vital information and interact with the first-order features to generate the interaction representation. Secondly, we adopt Euclidean distance to replace the traditional matrix transformation to capture the semantic similarity between aspect terms and context words. Finally, we form a joint representation to focus on the meaningful words in the sentence. We conduct extensive experiments and comparisons on SemEval 2014, SemEval 2016, and Twitter datasets. Experimental results demonstrate the effectiveness of our model.

OBJECTIVE: The goal of this study is to develop a robust Time Event Ontology (TEO), which can formally represent and reason both structured and unstructured temporal information. MATERIALS AND METHODS: Using our previous Clinical Narrative Temporal Relation Ontology 1.0 and 2.0 as a starting point, we redesigned concept primitives (clinical events and temporal expressions) and enriched temporal relations. Specifically, 2 sets of temporal relations (Allen's interval algebra and a novel suite of basic time relations) were used to specify qualitative temporal order relations, and a Temporal Relation Statement was designed to formalize quantitative temporal relations. Moreover, a variety of data properties were defined to represent diversified temporal expressions in clinical narratives. RESULTS: TEO has a rich set of classes and properties (object, data, and annotation). When evaluated with real electronic health record data from the Mayo Clinic, it could faithfully represent more than 95% of the temporal expressions. Its reasoning ability was further demonstrated on a sample drug adverse event report annotated with respect to TEO. The results showed that our Java-based TEO reasoner could answer a set of frequently asked time-related queries, demonstrating that TEO has a strong capability of reasoning complex temporal relations. CONCLUSION: TEO can support flexible temporal relation representation and reasoning. Our next step will be to apply TEO to the natural language processing field to facilitate automated temporal information annotation, extraction, and timeline reasoning to better support time-based clinical decision-making.

AI benchmarking becomes an increasingly important task. As suggested by many researchers, Intelligence Quotient (IQ) tests, which is widely regarded as one of the predominant benchmarks for measuring human intelligence, raises an interesting challenge for AI systems. For better solving IQ tests automatedly by machines, one needs to use, combine and advance many areas in AI including knowledge representation and reasoning, machine learning, natural language processing and image understanding. Also, automated IQ tests provides an ideal testbed for integrating symbolic and sub-symbolic approaches as both are found useful here. Hence, we argue that IQ tests, although not suitable for testing machine intelligence, provides an excellent benchmark for the current development of AI research. Nevertheless, most existing IQ test datasets are not comprehensive enough for this purpose. As a result, the conclusions obtained are not representative. To address this issue, we create IQ10k, a large-scale dataset that contains more than 10,000 IQ test questions. We also conduct a comparison study on IQ10k with a number of state-of-the-art approaches.

The study of paraconsistent reasoning with ontologies is especially important for the Semantic Web since knowledge is not always perfect within it. However, classical OWL reasoners cannot support reasoning with inconsistent ontologies. In this article, the authors present a plugin-based framework called prose to provide rich paraconsistent reasoning services for OWL ontologies, whose architecture contains the three following parts: a classical OWL reasoner, a multi-valued transformer, and an OWL API connecting with them. Within the proposed framework prose, they implement different multi-valued paraconsistent reasoning in the OWL. Moreover, they select three popular classical OWL reasoners and two typical kinds of reasoning services for users. As the authors excepted, prose does exactly enable current classical OWL reasoners to tolerate inconsistency in a simple and convenient way. Finally, they evaluate the three reasoners in a united framework (prose) and, as a result, those results can amend the analysis of the three reasoners on inconsistent ontologies.

In this paper, we analyze some basic features of SPARQL queries from practical world in a statistical way. In particular, we focus on three statistic features including the occurrence frequency of triple patterns, fragments, and well-designed patterns and four semantic features including monotonicity, non-monotonicity, weak monotonicity and satisfiability. All the features contribute to characterize SPARQL queries in different dimensions. We hope that this statistical analysis would provide some useful observations for researchers and engineers who are interested in what real-word SPARQL queries look like, so that they could develop some practical heuristics for processing SPARQL queries, as well as build SPARQL query processing engines and benchmarks. In addition, our research facilitates to reduce scope of the problems by avoiding some cases that may not occur in practice.